Autodesk’s documentation for the Revit Public MCP Server now lists three of them. The Read server handles “Querying, searching, and reporting on Revit models without making changes.” The Write server handles “Making controlled changes to Revit models using supported write tools.” The Experimental server carries everything in both plus more, and Autodesk lists its use as “Evaluating the newest capabilities and advanced workflows.”
The same page warns that the Write and Experimental servers can modify your model, that proposed changes should be reviewed, and that backing the model up first is worth considering. All three are “currently supported only on Autodesk Revit 2027.2 during the Technical Preview,” and Autodesk notes that “Features, supported tools, and behavior may change before the General Availability (GA) release.”
That is a materially different object from a read-only preview, which is what the server was when it arrived. The read-only server came first: Autodesk’s Japanese-language developer blog post of 29 September, by Toshiaki Isezaki, describes the update for the Revit 2027.2 release and states that alongside the existing Read-only MCP server, Write and Experimental servers are now provided as separate installers. AEC Magazine’s account is that the first official Revit MCP server launched in June as a read-only technical preview, which is consistent with Autodesk’s sequence, and that date is reported here as the magazine’s.
Officially supported write access from a language model into Revit is new — the server arrived read-only — but Autodesk is not alone in opening that path. Graphisoft’s Archicad MCP beta is described on its own product page as helping users “make updates, batch edits, and complete complex workflows” in the model. Either way, the records that would track such a change are a separate problem from the capability itself.
The Interview That Names the Problem
AEC Magazine published a Q&A on 2 November with Igor Tsinman, whom the magazine identifies as CEO of AMC Bridge, a software engineering services firm that builds integrations for and on top of CAD platforms. The piece is about what the magazine calls the home-brew estate: the growing stock of AI-built applications inside architecture and engineering firms that, on the magazine’s account, nobody formally owns.
Tsinman’s account of why that stock is growing is the uncontroversial part. AI lets the person who actually has the business problem, he says, “test an idea of the solution themselves,” without waiting for a conventional software project. His firm’s internal research, as reported in the interview, puts the time saving on routine coding and general logic at 70 to 80 percent — and he is explicit that the saving collapses for advanced geometry and specialised data, particularly 3D data processing, which still demands mathematical knowledge and careful verification. Those figures are AMC Bridge’s own and are reported here as such.
His argument is that the cheap part was never the expensive part. Most of the cost, on his account, sits in understanding the workflow and proving the application works against real project information — testing, security, integration and maintenance — none of which got cheaper. The prescription he offers is a slogan, and he says so: an established product may cover 80 percent of a requirement, with a firm’s own way of working living in the remaining 20 percent, so “The 80/20 rule remains a useful guide” and the approach is “Buy the foundation, build the workflow.”
The sharper observation is about categories, not costs:
“scripts are deterministic, while AI agents are not”
Firms have governed macros and scripts for decades, loosely, on the understanding that the same input produces the same output. Tsinman’s point is that an AI-driven change does not belong in that category and should not inherit its light-touch treatment. On approvals he is notably blunt about his own category: “I am not sure that we are at the point where we can trust an AI agent in the approval process.”
He applies the same caution to the protocol itself. MCP servers should write to production models only with clear permissions, validation and audit logs — and he says he would not make the MCP server “the foundation of a ten-year integration strategy.” That caution is worth weighing against his commercial position: per the interview, his firm sells an MCP Connector for Autodesk Platform Services and Procore.
Two of his answers are about where liability lands. On his own firm’s exposure, he says AMC Bridge stands behind what it agrees to deliver but “cannot assume responsibility for later modifications, uses outside the agreed scope.” On self-hosting, which is often offered as the answer to data-governance anxiety, he declines the comfort: “Keeping the model inside the firm does not guarantee that its output is correct.” His governance prescription is procedural — an amnesty for tools already built, an audit, a secure sandbox, possibly a temporary freeze on releasing unverified tools, and for anything shared, a named owner, company-held code and documentation.
One factual thread in the interview this publication could not stand up is the suggestion that a named large practice wrote its own Revit MCP server rather than wait for Autodesk’s. The claim appears in the interviewer’s question rather than in Tsinman’s answers, and searches found no public confirmation from the firm, so it is left out here rather than repeated.
The liability horizon the magazine raises is real, though it is worth stating more carefully than an interview question does. UK law firms’ published summaries of the Building Safety Act 2022 — for example Thrings’ note on the extended limitation periods — describe limitation for Defective Premises Act section 1 claims relating to new dwellings as extended from six years to 15 years prospectively and 30 years retrospectively, with section 135 of the Building Safety Act extending the period to 15 years for buildings completed on or after 28 June 2022. Those summaries differ on scope and on which related provisions are in force, and the Thrings note itself observes that the case law is limited and untested. The relevant point for this story does not depend on resolving that: the window in which someone may be asked what changed a model is measured in decades, and the audit trail for an agent-driven change is being designed now, or not at all.
What the Archive Already Shows
This is where the analysis is ours, and the pattern is visible across the not quite five months of coverage in this archive.
The supply side has moved fast. Revizto shipped an MCP server to open project data to enterprise AI. Graphisoft put MCP into an Archicad 30 beta. Nemetschek turned its Max tier into a cross-brand offering in which the MCP servers are the part that travels. Arcadis and Autodesk built a collaboration around MCP and the Autodesk Assistant. EarthCam exposed pre-analysed detections through connectors and MCP. Pirros wired skills into Revit for quality management. Every one of those is a vendor making it easier for a model to reach project data — and, in the Graphisoft and Nemetschek cases, to change it.
The governance side has produced far less. Aurigo certified an AI management system to ISO 42001 — the closest thing in this archive to a company answering the question with a standard rather than a feature. And the cost of getting document control wrong is not hypothetical: Turner Construction disclosed a data breach, and the ransomware group that claimed it says it also took engineering documents — a claim Turner will not address. Either way, the episode is a reminder that project models and drawing sets are a sensitive asset, not an incidental one.
Set those two lists side by side and the asymmetry is the story, with the comparison scoped to what this newsroom has published rather than to the industry as a whole. The archive is thick with vendors building the plumbing that lets agents read and now write to project models, and thin on the records that would let a firm answer, years later, which tool changed what and who authorised it. Tsinman’s list of what a firm should be able to show — which automated tool was used, where, who authorised it, which version ran, and what it changed — is not a technically difficult list. On the evidence of this archive, it is simply a list that far fewer companies have announced work on than have announced the connectors.
The reason to take a vendor interview seriously here is narrow, and it should be stated as a limit rather than an endorsement: AMC Bridge sells the governed-integration work it is recommending, and the Q&A carries a disclosure that HP and NVIDIA support AEC Magazine’s AI content while the coverage is editorially independent. The argument survives the conflict of interest because the underlying facts do not come from the vendor at all. They come from Autodesk’s own help page, which says in plain language that two of its three Revit MCP servers can change your model, and suggests considering a backup before they do.
